EVO-ICL Vault Prediction: A Data Wrangling Framework Integrating Multicenter Big Data and Machine Learning

The aim of this work is to predict the implantable Collamer lens (ICL) vault using machine learning (ML) algorithms and a data wrangling approach based on multicenter big data.

This retrospective cross-sectional study developed ML models using preoperative biometric data and ICL vault from 6715 eyes across five hospitals. Mutual information regression was employed to identify important parameters. A digital vault information system (DVIS) was constructed for data wrangling. ML models integrated with DVIS were used to develop ICL vault prediction and classification models, which were validated in both internal (6552 eyes) and external (163 eyes) validation.

The XGBoost model combined with DVIS exhibited statistically superior performance in ICL vault prediction, with lower mean absolute error (MAE) of 39.15 μm (internal validation) and 149.72 μm (external validation) compared to other ML algorithms. The R² value was 0.86 in the internal validation. For ICL vault classification, the XGBoost algorithm achieved accuracies of 81.4% (internal validation) and 57.27% (external validation), representing accuracy gains of 27.1% and 10.2% respectively, compared to traditional ML algorithms.

The development of DVIS is valuable for ICL vault prediction models, as it provides a data wrangling strategy that improves ML efficiency. Experimental results confirm the applicability of this synergistic method in enhancing existing ML approaches for ICL vault prediction, thereby facilitating more informed clinical decision-making in ICL implantation surgery.